(OK) Onyx - Structural - Equation - Modelling - 2018 PDF

An Introduction to Structural Equation Modelling (SEM)
Robin Beaumont 02/01/2018

An updated version to the introduction to SEM chapter in my book: Health science statistics using R and
R Commander by Robin Beaumont http://amzn.eu/g76aTP8
Books SEM webpage: http://robin-beaumont.co.uk/rbook/sem/index.html
Table of Contents
65 Structural Equation Modelling (SEM) ...................................................................................................... 2
65.1 The background to SEM ....................................................................................................................... 2

65.2 Path diagrams ...................................................................................................................................... 3
65.2.1 The model equations ................................................................................................................... 4
65.2.2 Interpreting a Path model (i.e. no circles) ................................................................................... 5
65.2.3 Interpreting a Standardised SEM model (i.e. circles and squares) .............................................. 6
65.2.4 Computing R squared from a regression using SEM .................................................................... 6
65.2.5 Calculating Direct, Indirect and Total Effects ............................................................................... 7
65.3 Latent variables .................................................................................................................................... 8
65.4 Covariances rather than raw data are used in model development ................................................... 9
65.5 The identification problem and sample size ...................................................................................... 10
65.6 Constraining parameters ................................................................................................................... 10
65.7 Model fit............................................................................................................................................. 11
65.7.1 RMSEA Root Mean Square Error of Approximation .................................................................. 12
65.8 Power and sample size determination .............................................................................................. 13
65.9 Why and When to use Structural Equation Modelling ...................................................................... 13
65.10 Stages of Structural Equation Modelling ........................................................................................... 14
65.11 A basic SEM model – the Hozinger & Swineford 1939 data analysed using Onyx (Ωnyx) ................ 14
65.11.1 Preliminaries .............................................................................................................................. 14
65.11.2 Creating an SEM model in Ωnyx................................................................................................. 15
65.11.3 Model estimation ....................................................................................................................... 19
65.11.4 The results .................................................................................................................................. 19
65.11.5 Obtaining the results.................................................................................................................. 20
65.12 Adding intercepts - Mean structures ................................................................................................. 21
65.13 Reloading a saved model ................................................................................................................... 22
65.14 Linking Ωnyx models to R ................................................................................................................... 22
65.14.1 Lavaan ........................................................................................................................................ 22
65.14.2 OpenMx (formally Mx) ............................................................................................................... 28
65.15 Structural Equation Modelling directly in R using the sem package ................................................. 28
65.16 A complex example ............................................................................................................................ 31
65.17 Extending SEM ................................................................................................................................... 32
65.18 Reviewing SEM research .................................................................................................................... 32
65.19 Tricks and Tips .................................................................................................................................... 34
65 Structural Equation Modelling (SEM)
Structural Equation modelling, SEM for short, allows you to develop and test models that
Self test questions
Whereas relatively few people consist of regressions, correlations and differences in means between groups. SEM is a
actually carry out SEM statistical technique that has developed from the concepts of covariance and correlation,
analyses a much large number
read books and articles therefore all the facts you know about correlation, including its limitations and pitfalls apply
reporting such analyses.
Therefore, for those of you
to SEM. Correlations themselves form the basis of path analysis (PA) and confirmatory factor
who only want to learn about analysis (CFA), which are forms of SEM. This will be explained diagrammatically on the next
SEM rather than carry out a
SEM analysis I have included few pages.
several self test questions in
this chapter.
65.1 The background to SEM

The statistical technique SEM attempts to minimise the difference between two sets of
X1 values (matrices of covariances=unstandardized correlations). One set is calculated from the
sample data while the other is generated from a hypothesised theoretical model defined by
Covariance or
Correlation
= standardised
the researcher. This idea can be expressed succinctly in the following equation.
covariance
the sample
X2
the difference between
and the model
min f ( S , ˆ )
Minimise
Multiple regression
X1 directed path
Paradoxically while this model comparison approach might appear very different
(single arrow)
from the traditional hypothesis testing approach (here you are hoping for the
error
Co-variation path least difference between the two) is can also incorporate the traditional approach
(double headed arrow)
(where you are hoping for a big a possible difference between the observed and
Y1
X2 zero effect models). For details see the online chapter SEM equivalent to basic
statistical procedures.
Box =
observed
X3 variable Schumacher and Lomax's excellent book, A Beginner's guide to Structural
Equation Modeling, 2010, provides details of what SEM is, from which I have
compiled the following list:
Factor analysis
1. Based on Correlations (Covariance)
2. Complex mathematical approach only made widely available with the use of suitable
observed
variables
paragrap software
also called
wordmean
3. Allows the definition of complex relationships using models (mathematically using
manifest PA1
variables covariance matrices which can be partially represented by diagrams)
sentence 4. Extends Regression (Path models)
cubes 5. Extends Confirmatory Factor Analysis (CFA)

PA2 6. Combines the two to form very complex models = Structural Equation Models (SEM)
visperc
7. Allows assessment of the degree to which a proposed model fits the sample data
lozenges
8. Allows comparison of models
latent variables
/factor/construct etc. There has been a long-ongoing debate as to what degree SEM models can help
establish causal rather than correlational relations (Bollen & Pearl, 2013). Possibly the
best approach is to consider the model as correlational and then re-interpret it in
terms of context to decide which are causal associations.
Let's take a look at the two building builds of SEM, Path modelling and (Confirmatory) factor
analysis (CFA) in a little more detail.
D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 2
65.2 Path diagrams
In SEM speak when the diagrams only contain observed variables they are called path
diagrams. The regression model on the previous page can be developed to
Path Diagram
X1 produce more complex configurations of observed variables where we no
longer just have several inputs (independent variables) along with a single
Y2 output (dependent variable), opposite there is additionally an intervening
(i.e. mediating) variable as well. Measured variables are also called manifest
Y1 variables. The Error represents a residual variance as the predictors only
X2
partially predict the value, think of it as a variable that mops up the
error
error inaccuracy in the prediction.
In contrast to the above regression and path models, models that contain both observed and
latent variables (discussed below) are called SEM models, such an example is the CPA model
above.
If the diagram has no numerical values on the lines it can either be a standardised or
unstandardized diagram, as in the above examples.
Each square represents a single observed variable that we specify. Traditionally the variables
would be of interval/ratio scale but software can now accommodate nominal/ordinal data.
The single arrowed lines are called [directed] paths and indicate that the variable at the
Standardized partial origin of the arrow has some influence on the target variable. These can be interpreted as
regression coefficients (unstandardized/standardised) partial regression coefficients or factor loadings depending
are also called beta
weights (ß)
where the lines appear on the diagram or if the diagram is unstandardized/standardised.
The double arrowed lines represent co-variation. If the values on the diagram are
standardised these lines represent standardised covariances = correlations or if the diagram
is unstandardized covariances. The multiple regression design shown on the previous page
shows correlations/covariances between the three inputs (predictors) (also see Kline, 2016
p.39).
Unfortunately, there are several different ways of drawing SEM diagrams and each SEM
software developer shows or ignores different aspects. A common aspect that is omitted is
the estimate of the variance of each observed/latent variable.
Path Diagram - Modified RAM format
X1 Because a variable's estimated variance is equivalent to the estimate
of its covariance with itself, i.e. cov(x,x) = var(x) we can use the
double-headed arrows again to represent this; they simply turn back
Y2 on themselves.
Y1
X2 The Modified RAM-format path diagram does this, (Fox, 2006 p.476;
Kline, 2016 p.130) as shown opposite. The error 'variables' now simply
1 become another co-variation arrow. Furthermore, including this visual
aspect aids interpretation of the model output, specifically the parameter
Creating SEM models estimates. You can now calculate exactly the number of parameters that are
going to be estimated by adding up all the single and double headed arrows
diagraming omitting those that have an identification constraint, distinguished by having a
tool value attached to them (e.g. y2 has a constraint of 1, in other words it is a fixed
value and not estimated). Eventually, whichever style of diagram you use, it will
matrix equations ultimately be converted into a set of matrices, which we will now briefly
(mathematical discuss. Alternatively, you can also specify a SEM model using various scripting
description)
languages which we will look at in the practical section in this chapter.
scripting
language

65.2.1 The model equations
There are two main ways of expressing the SEM model as a set of matrices. Firstly those
developed by Joreskog & Van Thillo, 1972 culminated in the development of the LISREL
software which specifies a set of matrixes by Greek letters
(http://web.pdx.edu/~newsomj/semclass/ho_lisrel%20notation.pdf). LISREL is still one of
the most popular commercial SEM software packages. Schumacker & Lomax, 2010 provide a
chapter describing the LISREL matrix approach
In 1980, McArdle suggested an alternative formulation, simplifying the 8 matrices needed by

LISREL to just 4 to create the Reticular action model (RAM). Fox, 2006 provides details of the
RAM approach along with how this is implemented in the R SEM package.
If you consider only those variables which have one or more single arrowed paths pointing
towards them, each group can be considered to be a separate equation (actually a
regression equation). In the diagram below, we have one such group.
Let's look at the 'predictors of mortality' regression example below produced using the free
Ωnyx program. When you right click on the diagram you can view either the RAM or LISREL
matrices (menu option below). I have shown the RAM option.
Predictors of mortality While it is not necessary to

Diagram shows both unstandardized and standardized understand the underlying matrices,
estimates in brackets
it is sometimes useful to inspect
them when either the model does
not produce the expected results, or
you are unable to obtain a result.
Variables:
Education, Popden, Nonwhite, Mortality
Means (defaults to zero)
m = ( 0.00 0.00 0.00 0.00 )

Asymmetric arrows
/ 0.00 0.00 0.00 0.00 \
A = | 0.00 0.00 0.00 0.00 |
| 0.00 0.00 0.00 0.00 |
\ Educ..y=25.51 Popd..y=0.01 Nonw..y=4.0 0.00 /
Covariance between variables Symmetric arrows
arrows
/VAR_n=0.70 COV_n=-286.04 COV_e=-1.55 0.00 \
S = |COV_n=-286.04 VAR_n=2108954.61 COV_n=-163.74 0.00 |
|COV_e=-1.55 COV_n=-163.74 VAR_e=78.26 0.00 |
\ 0.00 0.00 0.00 VAR_y=1541.64 /
Filter of observed
The four RAM matrices Var = 'error' variance for a variable variables
Defines the values of all the variables, constants, / 1 0 0 0 \
latents and their relations: F = | 0 1 0 0 |
• m = Means. By default all the values are set | 0 0 1 0 |
to zero the model using mean centred \ 0 0 0 1 /
scores.
• A = Asymmetrical arrows = regression

weights (single arrowed lines) The above matrices are manipulated to produce the various results
• S = Symmetrical arrows = the covariance
shown in a SEM output. Specifically, the S matrix of OBSERVED sample
matrix among the variables in the model. values are compared against those PREDICTED by the model stored in a
• F = Filter = index of variables in the model matrix called C, which is calculated from all the above matrices.
where, observed =1; constants and
latent's=0. There are only observed
variables in this model. (this matrix is also
called J by Fox, 2006)

The computer program then attempts to minimise the difference between the two sets of
values. This description leading to two questions:
1. What is meant by difference between?

There are various ways of measuring the difference between the two matrices and the
computed values are often called fit or discrepancy functions. Such functions go by the
name of ordinary least squares (OLS), asymptotically distribution free criterion (ADF),
generalised least squares (GLS) and maximum Likelihood (ML) where ML is the standard
technique in SEM. In ordinary least squares the result is the sum of the squared
difference between the two corresponding items in each matrix. In all these criteria the
minimum value is zero when both matrices are the same.
2. Is there any strategy to finding an improved fit between the two matrices each time
around?
By making use of one of the above computed discrepancy function values we can
compare these for difference guesses of our parameters and hopefully gradually home
in until we get a sufficiently small value. There are special search algorithms build into
SEM computer programs to help it gradually 'home in' but because the search program
may go widely astray we can usually also set a maximum number of times (i.e. iterations)
it can have a shot at finding a sufficiently close answer.
We will now look at a typical SEM diagram showing a set of model path estimates.
65.2.2 Interpreting a Path model (i.e. no circles)

The diagram below contains more information than the previous ones so I will discuss each
aspect separately:
Medical Model of quality of life for cardiac patients (Romney et al 1992) n=469 The results are taken from a correlational
Standardised Parameters study concerning 469 cardiac patients
0.31 0.925 where the investigators were keen to
Severity of Diminished
E5
Symptoms of Socioeconomic find out the associations between
Illness various measures.
0.48
0.14 SEM is usually carried out on large
0.18
0.838 samples and 469 is small compared to
Correlation
Low morale E1
= standardised covariance many studies. Also usually, the study is of
0.17 a retrospective/correlational design
0.54
although often there are attempts in the
Neurological Poor 0.842 discussion of such papers to interpret the
interpersonal E4
dysfunction
relationships
relationships as causal. We will return to
-0.13
this issue later.
Proportion of
Beta weights unexplained Title of Standardised Parameters - this
= standardised partial regression coefficients variance
indicates that the variables are
Reference: Romney D M. Jenkins C D. Bynner J M. 1992 A structural analysis of health-related standardised (i.e. z scores). This allows
quality of life dimensions. Human Relations, 45, 165-176 Discussed in
Grimm & Yarnold 1995 Reading and understanding multivariate statistics p.84. the various parameters in the diagram to
be compared and interpreted in a specific way described below.
Double arrow lines – These values indicate a correlation where the modelling process
estimates the value. Variables that do not have lines between them specify that the
correlation between them is set to zero in the model (i.e. they are independent).

Single arrow lines - These are beta weights (standardised partial regression coefficients)
which are identical to those in multiple regression. With this knowledge, it is possible to
interpret them as you would do in a regression equation. For example, a level of Severity of
symptoms of illness one full standard deviation above the mean predicts an increase in low
morale by 0.48 standard deviations above the mean controlling for Neurological dysfunction.
Error terms (E1, E4, E5)– This is the unexplained variance, that which is either measurement
error (random) or not explained by the model. This diagram uses the 'e' style to show them
rather than the self directed double arrows.
Proportion of unexplained variance values – these are also called the standardised residual
Self-test: variances (SRV). One minus these values gives the
2
1. Please complete this sentence. A level of Neurological R value (proportion of variance explained). So the
dysfunction one full standard deviation above the mean SRV x100 gives the % of the variance that is not
predicts an increase in poor interpersonal relationships by
____ Standard _____ below the mean controlling for predicted by the specific inputs that have arrows
Symptoms of illness. pointing to that variable. In some diagrams and
2. What does the value 0.54 represent between 'Low morale' and software (e.g. EQS) the square root value is given
'Poor relationships'?
instead.
3. What do the values from E5, E1 and E4 suggest about the
model? How do they relate to R2.
Direction of the single arrow - this does matter as it
4. What does the double pointed arrow ( <--> ) indicate. defines which variable will have the error term (the
5. There is no path between 'Neurological dysfunction' and 'Low variable that the arrow points to) and also how to
morale' in the model. What does this say about the model
specification? interpret the path coefficient.
6. Does the above model more closely represent a Regression
or Factor analysis? Give reasons for your answer. To check your understanding I have included some
7. How do you think you might calculate the correlation between
self-test questions opposite.
Severity of illness and Poor interpersonal relationships
(discussed further below). 65.2.3 Interpreting a Standardised SEM model
(i.e. circles and squares)
Interpreting an SEM model
SVR= containing both observed and
standardized
residual
latent variables (discussed latter) is
variance the same as for the above path
model.
65.2.4 Computing R squared

from a regression using SEM
Utilising the standardized path
Predictors of mortality coefficients in the SEM diagram
opposite along with the simple
correlations (below) we can obtain
R2 by using the equation discussed
in the multiple linear regression
chapter.
R 2  1 rYX1   2 rYX 2  3 rYX 3

 (.35)(.510)  (.19)(.261)  (.57)(.644)
 .1770741  .0488818  .3694264
 .595 (3 decimal places )

Alternatively, we could have simply calculated:
(1 - the standardised residual variance (SRV) of Mortality) = 1-.41 = 0.59.
I hope that this strengthens your belief that the SRV is a measure of what
remains unexplained after the regression analysis, as well as the idea that the
proportion of variance explained can be calculated from the beta weights and
the simple correlations between each input and the output.
To display the standardised path estimates alongside the unstandardised ones in
Ωnyx you need to select each path and then select the menu option shown
opposite.
65.2.5 Calculating Direct, Indirect and Total Effects
In a structural equation model it is
Standardised Estimates possible to calculate the various
(although you can also use unstandardized ones to work out the various effects) effects one or more variables have
Fitness upon another in the model via
-.260 Direct effect of Fitness upon illness = -0.260 other variables as well as directly.
From the example opposite it might
-.109 Indirect effect of Fitness upon illness = (-0.109)(0.291)
Illness
appear easy to calculate such
= -0.031719
effects, however this is not the case
Total effects are the sum of all direct and indirect with complex models as the indirect
effects of one variable on another
Stress
.291 path may follow tortuous routes. To
Total effect of Fitness upon illness = -.260 + -0.031719 help simplify this problem Sewell
Taken from Kline 2005p. 125 -129 adapted = -0.291719 Wright in the 1920's specified three
rules:
Admissible tracing rule example
example 1. You can trace backward up an arrow and then forward along the next, or forwards from
A one variable to another, but never forward and then back.
B
2. You can pass through each variable only once in a given chain of paths.
C
D 3. No more than one bi-directional arrow can be included in each path-chain.
Indirect effect of B to C = BAC admissible (rule 1)

Remember that the SEM model you have created will undoubtedly not contain
In contrast BDC is not admissible
all the possible paths between each variable so that there will also be
Taken from Loehlin 2004 p.9 (adapted)
unanalysed effects. Similarly, there may also be spurious effects in the model
where you create a path that goes against the flow of
Adapted from Grimm & Yarnold, 1995 p.84 the arrows at some point.
Standardised estimates E5
Diminished Once again, because you are more likely to need to
0.31 0.92
SS interpret and calculate these measures from a diagram
rather than carrying out a whole analysis yourself I have
Severity of
Symptoms of Illness
included an exercise below.
0.48 0.14
Self test:
E1 1. What is the total effect of Severity of Symptoms of
0.18 Low morale 0.84
Illness on Poor Interpersonal relationships?
0.17
2. What is the total effect of Neurological dysfunction
upon Diminished Socioeconomic Status?
Neurological 0.54 3. What is the direct effect of Diminished Socioeconomic

dysfunction Status on Poor interpersonal relationships?
Hint: in reality often pathways which have very small coefficients
-0.13 E4 (i.e. less than 0.1) are ignored.
Poor interpersonal
0.84
relationships

65.3 Latent variables
So far all the variables in the models we have discussed in detail only contain observed variables (boxes).
However we can also model latent variables. These are variables that are created from the data and mop up
variation in the model that is not directly taken into account by the directed (regression) paths. I can
demonstrate this
by considering a
regression model
which I have
depicted two
ways below.
In the first the

dependent
variable is
modelled with a
error variance
producing a
unstandardized
value of 1541.64.
(SD = 39.25)
Alternatively I
have also
modelled the
regression where
the dependant
variable has no
variance itself but is associated with a latent variable which I have called error. This latent variable consists
of the residuals scores, that is the difference between the expected and predicted score for each case. The
first table below shows the values for the first ten cases along with the residual and predicted (fitted) values
from a standard regression analysis. The second table shows the values from the error latent variable. You
can see that the residuals column from the traditional analysis is equivalent to this.
From traditional regression analysis:
rowname Education Popden Nonwhite Mortality fitted residuals
akronOH 11.4 3243 8.8 921.9 912.262883030003 9.63711696999752
albanyNY 11 4281 3.5 997.9 909.522843048244 88.3771569517563
allenPA 9.8 4260 0.8 962.4 929.164384657685 33.2356153423151
atlantGA 11.1 3125 27.1 982.3 992.17493049223 -9.87493049223009
baltimMD 9.6 6441 24.4 1071 1046.01268876674 24.9873112332572
birmhmAL 10.2 3325 38.5 1030 1062.32122750566 -32.3212275056638
bostonMA 12.1 4679 3.5 934.7 884.631050132788 50.068949867212
bridgeCT 10.6 2140 5.3 899.5 909.894939626843 -10.3949396268433
bufaloNY 10.5 6582 8.1 1002 958.979251244517 43.0207487554831
cantonOH 10.7 4213 6.7 912.3 929.433784297092 -17.1337842970919
From Onyx:
Education Popden Nonwhite Mortality error
Standardised scores Raw score
akronOH 0.4266666666666641 -623.0500000000002 -3.070000000000004 -18.481666666666456 9.637116969994196
albanyNY 0.02666666666666373 414.9499999999998 -8.370000000000005 57.518333333333544 88.37715695175669
allenPA -1.1733333333333356 393.9499999999998 -11.070000000000004 22.018333333333544 33.23561534231747
atlantGA 0.12666666666666337 -741.0500000000002 15.229999999999997 41.91833333333352 -9.874930492230668
baltimMD -1.3733333333333366 2574.95 12.529999999999994 130.61833333333357 24.987311233272635
birmhmAL -0.773333333333337 -541.0500000000002 26.629999999999995 89.61833333333357 -32.32122750565975
bostonMA 1.1266666666666634 812.9499999999998 -8.370000000000005 -5.681666666666388 50.06894986721164
bridgeCT -0.3733333333333366 -1726.0500000000002 -6.570000000000005 -40.88166666666643 -10.394939626850245
bufaloNY -0.47333333333333627 2715.95 -3.770000000000005 61.61833333333357 43.020748755494466
cantonOH -0.273333333333337 346.9499999999998 -5.170000000000004 -28.08166666666648 -17.13378429709057
Latent variables therefore provide a method of modelling concepts such as measurement

error and shared variance as we do in factor analysis and similar techniques. You can even
use this approach to model repeated measures, multilevel models and Meta-analysis
(Cheung 2015).

Another use of a latent variable is to add what is called a phantom
variable (cheung 2016 p.38). The variance of the phantom latent variable
is set to zero and adds nothing to the model. What is does is allow the
modeller to impose a constraint such as making is non-negative.
We will now see how it is not always necessary to use the raw data to
carry out a SEM analysis.
65.4 Covariances rather than raw data are used in model development
From the above discussion it should
now be clear that calculation of the
Sample covariance matrix:
sample data: V1, V2, V3, V4, V5, V6, various path values (coefficients) only
V1, V2, V3, V4, V5, V6, V1,
makes use of summary statistics (i.e.
Each has N V2,
observations V3, covariances or correlations). You may
V4,
V5,
remember in the chapter on factor
V6, analysis we saw how a matrix of
This dataset now only has 21 data items! (shaded cells)
=V(V+1)/2 = (6x7)/2 = 21 correlations along with a value
may be thousands
of values for these specifying the sample size could be
6 variables
but you only used instead of the raw data; the same
end up with 21
values! goes for SEM analysis.
Related to this is the fact that authors

are encouraged to publish the observed covariance matrix, to enable others to repeat at
least part of the analysis when presenting a SEM analysis.
Just because the 'raw data' is not used in the calculation of the path values, it should not be
assumed that sample size is not important, in contrast, it is extremely important. This is
because we always need to assess if the various estimated values are any different from
zero. This means it is necessary to either produce p-values or confidence intervals, both of
which rely upon the calculation of standard errors (unless we use bootstrapping techniques),
which itself is directly dependent upon the raw sample size. Furthermore, because SEM
models often have a substantial number of variables and paths, a large sample size (as a rule
of thumb, of at least 200) is needed to begin to make any sense of the data.

65.5 The identification problem and sample size
The identification aspect is basically asking, do we have
The Parameter-Monster enough data to allow us to develop a unique solution
Please save me to join the for our model. The following discussion, adapted from
degrees of freedom Schumacker & Lomax, 2010 p.63, provides an
p m explanation. Suppose our model has two values we
data data data data
data
need to estimate (i.e. free parameters) and one data
item. We can express this in the form of an equation:
X + Y = 10
If this is the case, X and Y can take on any number of
values (e.g. x=1;y=9; x=2.5 y=7.5 etc.). There is no unique combination.
Imagine that each time the computer runs a program to find a solution to this
 under-identified dfM < 0 problem it just randomly assigns a value between 0.1 and 9.9 to X and therefore
not always - see Kline, 2011 provides a different solution for Y each time. When it is impossible to define a
unique value to each of our parameters in our model the model is said to be
under-identified and is useless. When we talk about identification, we are talking
 just-identified dfM = 0
about the degrees of freedom for the model (dfm) - in other words the amount of
data that can move freely after accounting for the parameter estimates. A under-
identified model possesses negative degrees of freedom and no unique solution
☺ over-identified dfM >0
can be found.
In contrast, the desirable situation is to have a model which is preferably over-identified or,
failing that, just identified. Ideally we want a model that estimates the smallest number of
parameters but at the same time provides the best fit, thus maximising the accuracy of the
estimates of our parameters (or more traditionally we say maximising the power of the
study).
This ratio of the of sample size to number of parameters to be estimated in an SEM is

important. Quoting Bagozzi & Yi, 2012 p.29: As Bentler and Chou (1987, p. 91) state that the
ratio “may be able to go as low as 5:1 under normal and elliptical theory, especially when
there are many indicators of latent variables and the associated factor loadings are large”.
But they also believe that “a ratio of at least 10:1 may be more appropriate for arbitrary
distributions.”. I will revisit this issue when discussing model fit indices.
65.6 Constraining parameters

One solution to the identification problem is to set one or more of the parameters to a
constant value, which is then known as a fixed parameter. An
Types of constrained parameters (Kline 2016 p. 129) example of fixed parameters are all the hidden paths in the model
1. Equality – two or more paths are set to equal one indicating a correlation fixed to zero. Looking back at the model
another
2. Inequality – this is where a value is specified to matrices section you can see these hidden values are represented by
be within a specified range the zeros. Anther example of this is known as an identification
3. Cross-group – This is where two of more paths
constraint. We saw this in the factor models chapter where one of
across independent groups in a model are set to
be equal. the loadings was set to 1 for each of the factors.
4. Proportional – one parameter is proportional to
another path estimate A constrained parameter is one that does not equal a constant but
5. Nonlinear – This is where one parameter is a
nonlinear relation to another such as its square.
has some type of restriction imposed upon it. Kline 2016 p.129
mentions five types listed opposite. A good example of one of these
can be found in my youtube video of the two independent samples t test as a SEM model
https://www.youtube.com/watch?v=Hak-fyDc8fg

I have provided a graphical summary of some of the aspects discussed in the above sections,
which you can use as a revision add.
have have
Variables Paths Values "parameters"
Fixed
Manifest Latent Directed Bi-directional Constrained Free
"constant"
"observed/measured" "factor/construct" "regression paths" "covariances/correlation"
equality
"standard" s
Between variables inequality
phantom variable
Cheung 2015 p.39 Within variable cross-group
error/residual variance variance (independ. vars) proportional
error/residual variance nonlinear
equivalent (depend. vars)
Cheung 2015 p.40
Along with the identification problem comes the problem of assessing how well overall the
model fits the observed data.
65.7 Model fit

There are many problems with taking an often complex model and obtaining a single value
from it to represent the idea of fit. As there is no acceptable solution to this problem every
SEM program provides over two dozen fit measures, and most research papers offer a
selection! Because single fit measures provide an overall averaged value they unfortunately
stop researchers looking at each individual parameter and its residual. There is always the
possibility that the model may fit very well for the majority of the parameters but then have
a large residual for one. If this is the case the researcher would then need to consider if the
particular variable was correctly specified in the model and also how important is was in
comparison to the other variables. I describe this as the does my bum look big in this dress
aspect. Because of this I recommend that you always start by looking at the standardised
residual matrix (Loehlin, 2004 p.69) which will highlight particular hotspots.
Schumacker & Lomax then suggest a three stage strategy (2010 p. 74):
1. Inspect two overall fit measures:
o Chi-squared - you are aiming for an insignificant result i.e. high p-value
(usually above .95) resulting from a small actual value. This value is
problematic both when you have df=0 or a large initial sample (number of
subjects, not the number of covariances). Various versions of this statistic
such as GFI and AGFI have been developed to take these problems into
account.
o RMSEA (Root Mean Square Error of Approximation) - is at present the most

popular measure, you can use it to produce confidence intervals or an
associated p-value.
2. Inspect significance of individual parameters. The critical ratio (CR) for each
parameter estimate is computed by dividing the estimate by its standard error. In
most SEM programs it is compared to a z or t value of 1.96 at the .05 level of
significance. Some software (i.e. EQS) indicate parameters that are statistically
significant by placing an asterisk (*) beside them. Often in the model development
process insignificant paths are removed from the model.

3. Inspect magnitude and direction of parameter estimates - does the value make
sense in the proposed theoretical model? Has the computer provided impossible
values such as negative variances (called Heywood cases) or correlations >1.
A typical set of model fitting statistics is shown

opposite for the morality model obtained from the
free Ωnyx software.
The chi-squared actual value is very small (E-13)

meaning that we would have a p-value approaching
1, Indicating a good fit. The AIC and BIC measures
are used to compare models where smaller values
indicate a better fit.
Overall model fit metrics – Be wary Baggozzi, 2012 quoting various writers recommends the following standards for
indications of a good fit! assessing adequate fitness of SEM models: CFI≥0.95, and SRMR≤.08, also see
• Chi-squared p-value ≥ .95 Schumacker & Lomax, 2010. Again both the values reported opposite on the
• CFI≥0.95 previous page indicate a good model fit. I'll discuss the RMSEA next.
• SRMR≤.08
• RMSEA ≤.10 "good" ≤ .05
"very good" fit
65.7.1 RMSEA Root Mean Square Error of Approximation

The RMSEA is the golden boy of fit indices having many attractive attributes. It uses a value
d, based the non-centrality parameter (Lambda =λ) of the chi-squared (2) distribution
which is simply the obtained chi-squared value of the model (i.e. Lambda =λ = obtained chi-
squared).
d
RMSEA 
df Taking the mortality example shown earlier we have a 2 = approximately zero. With 4 df in
a sample of 60 cases, producing a d value of (0 - 4)/59 = -.066, and a RMSEA equal to
where
√(.066/59), or -.001186. Because this value is less than zero we set the RMSEA value to zero.
The maximum value RMSEA can obtain is 1.
 2  df
d
N 1 The RMSEA measure was developed by Steiger (1998) who considers values below .10
"good" and below .05 indicative of "very good" fit. It is also possible to obtain a confidence
limit for RMSEA using the noncentral 2, distribution, described in the confidence intervals
for effect sizes chapter. The development of the RMSEA stems from Steiger's (Steiger, 1989)
belief that the appropriate question for statistical analysis in SEM is not assessing how
perfect the fit is but rather, the three following questions.
1. How well does my model fit my statistical population?
2. How precisely have we determined population fit from our sample data?
3. Does the fit still appear good when we take into account the complexity of the model and
its number of free parameters?
By developing a confidence interval for the RMSEA we can take into account the three above
issues testing a null hypothesis of poor, instead of perfect, fit. If the upper limit of the 90%
confidence interval lies above the desired cutoff, for example .10, we can then reject the
hypothesis that the fit of the model in the population is worse than our desired level.
Similarly, if the CI does not extend beyond this cut-off we can conclude our model is of
adequate fit. The width of the confidence interval will also provide information about the
accuracy of the estimate which is always useful. Because the RMSEA is based upon the non-
central chi-squared distribution it can also form the basis of a power analysis discussed next.
65.8 Power and sample size determination

The previous section described the general rule of the ratio of the sample size to the number
of parameters. However this rather questionably naïve approach has been improved upon
by several writters.
The easiest approach is to consider the RMSEA for two models from which you can readily
calculate the power. Schoemann, Preacher & Coffman, 2010 provide an online calculator
along with several other comparable measures, along with R code. Similarly the R package
SEM Tools contains a function plotRMSEApower() which allows the plotting of power across
a range of sample sizes much like Gpower. The investigation of power during the planning
stages of a study that intends to use SEM techniques can be a long and arduous process as it
requires not least the investigators to define their proposed models at the start of the
process and also make a guess (quite literally) of what the RMSEA value might be.
Therefore detailed power analysis is not for the faint-hearted and anyone planning to
undertake this I would recommend they start by reviewing the literature. I would advise
MacCallum, Browne & Sugawara, 1996; Hancock & Freeman, 2001; Lee, Cai & MacCallum,
2012 and Miles, 2003. Professor David Kaplan, University of Wisconsin Madison, provides a
brief review of these at: http://www2.gsu.edu/~mkteer/power.html
Rex Kline who has written a popular introduction to SEM (Kline, 2016) and provides a set of
notes concerning power estimation for a variety of SEM models, including some R code, at
http://psychology.concordia.ca/fac/kline/SEM/qicss2/qicss2setA.pdf
Professor Timo Gnambs at the Leibniz Institute for Educational Trajectories in Bamberg
provides an online R code generator for SEM power analysis at:
http://timo.gnambs.at/en/scripts/powerforsem
65.9 Why and When to use Structural Equation Modelling

The following is a list of the common reasons given for using SEM techniques:
1. Confirm Relationships / hypotheses

2. Test complexity (multivariate)
3. Longitudinal / Multiple groups (panel designs)
4. Test Specific relationships
5. Multiple observed variables
6. Take measurement error into account
To use Structural Equation Modelling successfully often requires time to dedicate oneself to
the SEM literature and also specific SEM software with the result that frequently those
researchers that become competent tend to apply it to almost all problems. Because of the
paradox between the complexity of the mathematics and interpretation of the results
compared to the relative ease of being able to draw the diagrams for a structural equation
model it is often inappropriately used.
One common problem often encountered with developing SEM models is that of
identification, which is discussed next.

65.10 Stages of Structural Equation Modelling
After having discussed identification issues, model fit indices and
SEM modelling process power analysis it is clear that the development of SEM models is
far more complex than just drawing a diagram. Complex models
How was the Where did it
knowledge originate? also require the analyst to know the research area in depth and
structured?
be able to make sensible informed decisions concerning
Basis of prospective models.
model Why am I
Do I have the
skills to use planning on
particular using this
software technique?
package?
Once you have a working model that the program is able to
SEM? estimate then you often need to change it to perform
code or
visual?
comparisons with various other models. Often only after you
have done this several times do you finally come to a decision
Shall I built Build the deciding which is the 'best' possible one.
alternatives
model
and compare
them?
I'll now demonstrate how you can create a simple SEM model by
Type of model Evaluate
Relationships? the model repeating the analysis described in the factor analysis chapter. I’ll
Variables / types
Parameters fixed
also demonstrate a more complex model.
/ free
Report
65.11 A basic SEM model – the Hozinger & Swineford 1939 data analysed using Onyx (Ωnyx)
Details of the Hozinger & Swineford 1939 dataset were discussed in the factor analysis
chapter. In this section we will use a separate program that produces R code called Ωnyx
(Ωnyx) to carry out the SEM analysis. This allows us to specify the model using a set of
graphical symbols which you may recognise from the factor analysis chapter.
65.11.1 Preliminaries
Watch my youtube video demonstrating the steps described below:
https://youtu.be/GN-HOWd31Ao (this link is case sensitive, the last character is lower case 'o' as in 'on')
You first need to download the dataset (tab-delimited format) from the link below and save
it locally: http://www.robin-beaumont.co.uk/virtualclassroom/book2data/grnt_fem.dat you
can also find it at the books website.
If you go to the link above the page

opposite will appear showing the
data, while on the page right mouse
click and save the data locally.

Before you can use the Ωnyx program you need to
have Java installed which you can find at:
http://www.java.com/en/
Then download Ωnyx from:
http://onyx.brandmaier.de/download/
The Ωnyx program, being a java file, has a jar

extension, which you save to a local folder.
As shown opposite, I created a folder on my d: drive called onyx.
To run Ωnyx Double click on the jar file. This brings up the main window of Ωnyx.
65.11.2 Creating an SEM model in

Ωnyx
The main window of Ωnyx contains a tip
of the day box which you can close by
clicking on the close button. The tip also
appears on the bottom bar of the Ωnyx
window.
The first thing you need to do is create a

new model. You can do this either by selecting
the option from the dropdown menu or by
double left mouse clicking in the desktop. The
model panel now appears. You can give the
model a name by right
mouse clicking
anywhere in the
window. I have called
it grnt_fem_m1.
model panel

We can now proceed to specify the model by either drawing the
symbols and then linking them to the fields in your dataset or, the
technique we are going to use now, loading the dataset into Ωnyx
dataset panel and dragging the fields across to the model panel.
To load the dataset into Ωnyx, either select the following menu
option or simply drag the data file from the file browser window.
File->load data
A hexagonal window (called the dataset panel) appears which

represents your dataset.
Ωnyx has context-sensitive popup menus, which means that if you

hold the mouse cursor over various elements you get information
about them
To add one or more variables to your model window simply drag them
from the dataset panel to the model panel. You can select more than
one using either the shift (continuous blocks) or CTRL keys
(discontinuous blocks).
Ωnyx gives each graphical representation of the variable the same name
as the variable in the dataset.
You can move and reshape the elements by selecting and then clicking
and dragging.
Similarly, you can move the model panel by dragging its top or bottom border (a hand
appears) or resize it by dragging the right hand lower corner .
We are going to create a type of SEM model known as a Confirmatory Factor Analysis (CFA)
and to interpret the findings of such an analysis is it best to have the variables standardised.
Luckily, you don't need to convert the variables as Ωnyx can do it for you.
variance
The error variance for each
variable is often simply labelled
model panel E1 to En where n is the
dataset panel number of variables in the
selection
indicators model requiring error
variances.
panel
resize area

To edit them we follow a similar process to that above. First select
the error variance (left mouse click), that is the little loop attached
to each variable then bring up the menu (right mouse click). This
menu allows you to change the path name, where I have
gone through each in turn and renamed them e1
to e6.
We now need to add the two latent variables

(called factors in the last chapter).
Either double right mouse click in an empty part of the

model panel
or
Right mouse click on the model panel to bring up the

menu shown opposite, then select the following:
Create Variable->Latent
We select each latent variable (factor) in turn and then bring up

the menu (right mouse click) to change the names to visual and
reading respectively.
We now need to fix

both the error
variances associated
with these latent
variables . Select the variance paths (left mouse click)
and then bring up the menu (right mouse click). Each
time select the Fix Parameter option from the menu.

2. right click and drag The last thing to do is to specify the relationships between the variables in the
to destination
model. First the paths between the observed variables and the latent
variables.
1. select origin
of path (left
click) To draw a path select the variable for the paths origin (left mouse click)
then:
Right mouse click on it and drag to the target variable.
I repeat this several times to create the model shown on the left.
Important: note that the arrows go from the latent variables to the
observed ones.
The style described above is the traditional way of defining a factor

model using what is known as reflective measurement. An alternative,
more contentious approach is to have the arrows pointing in the other
direction producing a formative measurement model. However there are
questions concerning the validity of this latter approach and I suggest
that anyone who believes otherwise consult various articles in
Psychological Methods, 2007, 12(2), specifically Howell, Breivik, &
Wilcox, 2007.
The only other path to add is that for the correlation (i.e. covariance)
between the two latent variables.
Select one of the latent variables by left mouse clicking on it and hold
down the shift key + right mouse click then drag to target
variable.
The far left screenshot shows that by hovering the mouse

cursor over this line we discover it is a fixed parameter
estimate whereas we want the model to estimate its
values. To change it:
Select the path (left mouse click) then right mouse click on
it to bring up the menu and select the following menu
option:
Free Parameter
You need to carry out this process for all the paths
directed from the observed variables to the latent ones as
well. Once again, you can check to see which ones are
fixed or allowed to be estimated by hovering over each
path.

65.11.3 Model estimation
While we have been specifying the model, the Ωnyx program has
been calculating the parameter estimates along with several model
fit values.
You can update the diagram to show the 'best' parameter
estimates by right mouse clicking on an empty part of the model
panel and then selecting
Estimation-> Show Best ML Estimate
65.11.3.1 Displaying standardised values

It is also useful to see the standardised estimates.
These appear in parentheses when you hover over
a path.
standardised estimate
To display the standardised path estimates alongside the

unstandardized ones in an Ωnyx diagram you need to select each
path and then select the menu option shown opposite. You can
select multiple paths by using the shift+ left mouse click combination.
65.11.4 The results

I have discussed in detail in the previous chapterl the meaning of the various results
from a Exploratory Factor Analysis, and the same interpretations apply here.
Factor loadings These are the paths coefficients between a

specific observed variable ('indicator') and a specific factor.
Higher values mean a closer relationship. They are equivalent
to standardised regression coefficients (β weights) in multiple
regression. Higher the value the better.
Communality (h2) This is the total influence on a single

observed variable from all the factors associated with it. It is
equal to the sum of all the squared [standardised] factor
loadings for all the factors related to the observed variable. The value ranges from
zero to 1 where 1 indicates that the variable can be fully defined by the factors and
has no uniqueness. In contrast a value of 0 indicates that the variable cannot be
predicted at all from any of the factors. The communality (h2) can be derived for each
variable by taking the sum of the squared factor loadings for each of the factors associated
with the variable. As these values are equivalent to the R squared values in multiple
regression, we can interpret them the same way; they represent the % of variability taken
into account by the model. Because we are hoping that the observed dataset is reflected in
the model we want these values to be as high as possible, nearer to one the better.
Uniqueness (u2) for each observed variable this is that portion of the variable that cannot be
predicted from the other variables (i.e. the latent variables). As the communality can be
interpreted as the % of the variability that is predicted by the model we can say that
uniqueness is the % of variability in a specific observed variable that is NOT predicted by the
model. This means that we want this value for each observed variable to be as low as
possible. Lower the better.

65.11.5 Obtaining the results
In Ωnyx the results are obtained by right mouse clicking on an empty part of the model panel
and then selecting Estimate summary.
u2 = uniqueness (=1-h2)
Standardised value=
Factor loading
When single line from
indicator squared value
same as commonality h2
all variables centred
estimate for each parameter
We say a factor loads on one

or more items/indicators
model fit indices
The results are shown above. It is important to realise that Ωnyx calculates the results 200
times and the final part of the estimate summary provides details of the results from which
we can see that luckily the results stabilised.
Overall model fit metrics – Be wary Because the above model is very similar to that presented in the factor
indications of a good fit! analysis chapter, mimicking most closely the Promax rotation result, we can
• Chi-squared p-value ≥ .95 interpret the paths between the latent variables and the observed variables as
• CFI≥0.95 commonality values and the error terms as being the measurement errors. The
• SRMR≤.08
slight differences between these values and those in the factor analysis
• RMSEA ≤.10 "good" ≤ .05
"very good" fit chapter, where u2i + h2i =1, is because here we have constrained several of
the paths to zero (i.e. there is no path between lozenges and verbal). This is
known as a Confirmatory Factor Analysis (CFA) approach in contrast to the previous chapter
where we followed what is known as an Exploratory Factor (EFA) approach.

As mentioned above when there is a single path from a observed variable (i.e. indicator) to
the factor (i.e. latent variable) the
communality is equal to the squared factor
loading which is the same as the squared
standardised path coefficient.
In the example opposite I have now added

what is known as cross factor loading where
both factors load on the Lozenges indicator.
Here the communality for lozenges is now
0.662 + .122 = 0.45 which is different from the
previous value of 742 = 0.5476 in other words
this is a less desirable value than we
previously had. In the section looking at the
lavaan output we will look in more depth at
the two models.
65.12 Adding intercepts - Mean

structures
For more complex SEM models, specifically
where we wish to investigate possible
differences between means in several groups, it is necessary to add this information to
the model. This is because so far we have been working with centred variables (i.e.
mean = zero), the so-called saturated means model.
Graphically to explicitly specify the means in a SEM model we add a special graphical
symbol, a triangle, to the model and then draw lines from it, pointing to those variables
for which you want the means to be included in the model. Adding this symbol implicitly
sets the menu option:
Mean Structure -> Explicit Means.
For further details see the online chapters at the books website:
An Introduction to Regression Modelling using Structural Equation Modelling
(SEM)
and
SEM equivalent to basic statistical procedures

65.13 Reloading a saved model
Once you have saved a model in Ωnyx when
you reload it all the connections to the data
will have disappeared. To reform these
connections you simply open the data file
again, then drag the appropriate field name
into the correct object in the model panel.
Repeat this process for each field.
65.14 Linking Ωnyx models to R

Ωnyx is designed primarily as a pedagogical tool allowing you to specify models graphically
which can then be run using actual R packages. Specifically it supports three R packages
lavaan, OpenMx and SEM. The way it supports these R packages is by producing the
appropriate R code for each. We will begin by looking at the Lavaan code.
65.14.1 Lavaan
I will demonstrate the use of Lavaan using the code generated by Ωnyx for the
Confirmatory Factor Analysis (CFA) described above. You first need to obtain
and load the data to go with the model into R. Copy the following code into
the R Console window.
hozdata<- read.table("http://www.robin-beaumont.co.uk/virtualclassroom/book2data/grnt_fem.dat",
sep="\t", header=TRUE); names(hozdata)
Within R you first need to install and load the

lavaan package:
the lavaan code
install.packages("lavaan", dependencies =
produced by Ωnyx
TRUE); library(lavaan)
The Lavaan package makes use of what it calls

lavaan model syntax which describes a latent
variable model, and luckily Ωnyx produces this
code for us. You can see this syntax by
selecting the lavaan option in the menu
shown above, this appears when you right
mouse click in an empty area of the model
pane. The resulting lavaan code window is
shown opposite.

Since the book was written the code generated has become
Lavaan code generated (2018)
library(lavaan); more verbose and includes all the path labels and an additional
modelData <- read.table(DATAFILENAME, header = TRUE) ; observed means section. The path labels appear just before
model<-"
! regressions the '*' symbol and in this context they should be seen as a way
visual=~visual_TO_cubes*cubes
visual=~visual_TO_lozenges*lozenges
of identifying modifiers not multiplication. The lavaan website
reading=~reading_TO_paragrap*paragrap provides further details
reading=~reading_TO_sentence*sentence
reading=~reading_TO_wordmean*wordmean http://lavaan.ugent.be/tutorial/syntax2.html
visual=~visual_TO_visperc*visperc
! residuals, variances and covariances As with R Commander this auto-generated code is more
visperc ~~ VAR_visperc*visperc A 'label'
cubes ~~ VAR_cubes*cubes indicated by complex than need be, these automated tools lack the
lozenges ~~ VAR_lozenges*lozenges the * to the aesthetics and creative originality of a human! But it does
paragrap ~~ VAR_paragrap*paragrap right of it
sentence ~~ VAR_sentence*sentence provide a very useful starting point.
wordmean ~~ VAR_wordmean*wordmean
visual ~~ 1.0*visual
reading ~~ 1.0*reading
The code is punctuated by several comments lines (!) where
reading ~~ COV_reading_visual*visual A mean each subsection defines a particular aspect of the model. I
! observed means structure
visperc~1; have reproduced the code below with detailed comments. To
appears to be
cubes~1; added! obtain the equivalent analysis to Ωnyx in R using the lavaan
lozenges~1; If you delete the mean
paragrap~1; structure you MUST package requires slight modification of the code which we will
sentence~1; remove the missing=FIML
wordmean~1;
now undertake. You can safely remove the path labels, the
option as well
"; observed means section and the missing=FIML option, as I
result<-lavaan(model, data=modelData, fixed.x=FALSE,
missing="FIML"); have done below:
summary(result, fit.measures=TRUE);
specify the data filename (text file) you wish to use for the model –by editing this value
or remove this line. I usually remove it. See below
load the lavaan package, we have already
done that, so you can remove this line
library(lavaan);
dat <- read.table(DATAFILENAME, header =
" or ' TRUE) ;
we now begin to specify the lavaan model, model<-" a comment line, you could use the # symbol instead
basically we have model_name <-" ! regressions
note the double quote that is important
visual=~cubes
visual=~lozenges =~
Latent variable (continuous variables) definitions.
reading=~paragrap The name of the latent variable is on the left of the "=~"
~~
reading=~sentence operator, while the terms on the right, separated by "+"
Variances and covariances. The "~~"
(‘double tilde’) operator specifies reading=~wordmean operators, are the indicators of the latent variable.
(residual) variances of an visual=~visperc Alternatively, as here each is on a separate line.
observed or latent variable, or a set of Read "=~" as “is measured by”
! residuals, variances and covariances
covariances between one variable, and
visperc ~~ visperc
several other
cubes ~~ cubes variances
variables. Several variables, separated by
"+" operators can appear on the right. This lozenges ~~ lozenges
way, several pairwise (co)variances paragrap ~~ paragrap
involving the same left-hand variable "*" are referred to as modifiers. 1 * sets the variance
sentence ~~ sentence to one, it becoming a fixed parameter
can be expressed in a single expression.
The distinction between variances and wordmean ~~ wordmean
residual variances visual ~~ 1.0*visual
covariances - only one in this model
is made automatically. reading ~~ 1.0*reading
visual ~~ reading
you need to specify the dataframe here. Change
"; "data=data" to "data = hozdata"
" or ' result<-lavaan(model, data=dat,
To indicate we have finished specifying now send the model definition to the lavaan function
the lavaan model. Followed either by a
fixed.x=FALSE);
SEMi colon (;) or a blank line summary(result, fit.measures=TRUE);
produce a set of summary results for our model
including fit indices
~ (not used in above model)
Regression definitions. The output (dependent) variable is on
the left of a "~" operator and the input (independent)
variables, separated by "+" operators, are on the right.
Taking into account the annotations described about we now have:

library(lavaan);
dat <- hozdata
model<-"
! regressions
visual=~cubes
visual=~lozenges
visual=~visperc
reading=~paragrap
reading=~sentence
reading=~wordmean
! residuals, variances and covariances
visperc ~~ visperc
cubes ~~ cubes
lozenges ~~ lozenges
paragrap ~~ paragrap
sentence ~~ sentence
wordmean ~~ wordmean
visual ~~ 1.0*visual
reading ~~ 1.0*reading
visual ~~ reading
";
result<-lavaan(model, data=dat, fixed.x=FALSE);
summary(result, fit.measures=TRUE);
The above summary function produces the output on the left

showing a range of model fit indices. To gain R squared and
standardised values you can edit the summary function to:
summary(result, fit.measures=TRUE, standardized=TRUE, rsq=TRUE)
To gain further information about the parameter estimates we use the parameterEstimates()
function which also provides confidence intervals, given below. I have shaded out two
irrelevant columns. standardized estimate
parameterEstimates(result, standardized = TRUE) either a standardised path
value, covariance or a
variance
I have also added below

some of the results from
the model with the cross
factor loading where both
factors loaded on the
Lozenges indicator. As
would be expected the
path from the Reading
factor to is statistically
insignificant indicating that
indicating regression weight or standard p-value given that it 95% ci
variance/covariance equals zero in the
its population value is equal to zero.
estimated error
value population
Unfortunately lavaan does not automatically produce
indirect/total effect values but the following tutorial explains how you can do it
https://www.youtube.com/watch?v=-B37sK9NTfI

65.14.1.1 Modification indices (MI)
There are been several attempts to automate both model creation and model modification.
Unfortunately both approaches can result in nonsense models where the optimum fitting
model (usually several) does not reflect a sensible theory in the particular research area. As
with factor analysis, SEM modelling is a mixture of both art and science.
Modification measures come in two main varieties:
1. Adding paths - Langrange multiplier tests (LM) provide information about how much
better the model would fit (i.e. 2 statistic would be reduced) if a particular
parameter were added to the model. That is if the parameter were allowed to be
estimated rather than being constrained/fixed.
2. Removing paths - Wald tests (backward search) indicate which parameters can be
dropped without affecting model fit.
The lavaan package takes the first approach with the modificationIndices() function, using
the output from our model modificationIndices(result) the output of which is shown below.
Expected Parameter Change (EPC).
improvement in model fit if Estimated parameter value if it were
parameter added added to the model
standardized EPC value While blindly looking through the
indices and purely deciding on
'improving the model' by adding one or two paths
with the highest mi values might be tempting this is
not really the best strategy as the modelling process
should also be informed by knowledge of the research
area.
There are no standard rules concerning a specific cut

off value of the mi to indicate necessary inclusion in
the model. Therefore it is often more sensible to only
display those values where the change is greater than
a specific value, e.g. 10, which can be achieved with
the following code.
change_results <-modificationIndices(result)
subset(change_results, mi > 10)
Looking at the results opposite, we see that the

greatest improvement in model fit would be achieved
by adding a covariance between visperc and
wordmean. Whether this makes sense in terms of the
theory of verbal and spatial intelligence would need
knowledge of the research area.

65.14.1.2 Residual analysis – analysing local fit
As mentioned earlier it is all well and good obtaining overall
(i.e. global) model fit measures but, as is the case with the
simple contingency table chi-squared introduced many
chapters previously, often the desire is to see where the
actual model fits best and least well. As we did in that
chapter, a residual analysis of the values, this time the
covariance residual matrix, provides such information.
In effect, each value in the residual matrix reflects the

difference between the observed and modal covariance
matrix. You can obtain three sets of values, raw difference,
standardised scores (which are similar to z scores), and
normalised scores which are more difficult to interpret (see
Bollen, 1989 p.259).
resid(result, type="raw")
resid(result, type="normalized")
resid(result, type="standardized")
# like z scores
If we look at the standardized residuals, we can see that the

largest is for the covariance between cubes and sentence
which misfits by around three
quarters of a standard deviation.
Given that none of them is greater
than 1.96 reinforces the belief that
this is not only a good model fit at the global level (from the
RMSEA value) but also a good model fit from the individual
parameter level as well.
65.14.1.3 Comparing models

Ωnyx provides an interesting
graphical approach to model
comparison.
You start by defining two, or

more, models. An easy way to
do this is to clone one from the
other, achieved by selecting the
Edit->Clone Model popup
menu option which appears
when you right mouse click
within a model panel. In the
example below I have cloned the
model then changed the path
from visual_to_lozenges to
reading_to_lozenges.

To compare these models, where they both have the same number of parameters, we
simply bring up the estimate summary box for both and compare the various fit indices. As
shown below, the left hand side one has a smaller RMSEA value (☺) and smaller AIC and BIC
values (☺) all indicating that it
is a better fit.
To compare models where in

one of them one or more of
the paths is constrained to a
particular value (called a
nested model = number of
estimated parameters is
different for each model) we
use a likelihood ratio test, as
discussed in the Logistic
Regression chapter. To obtain
this value in Ωnyx involves
using a unique method, we
simply right mouse click on
one model and then drag to
the comparator model. In the
example below, I have
constrained the covariance between the
two latent variables to 0.1 and then
compared the model to that where the
covariance is estimated. Hovering over the
black dot informs us that we have a -2log
likelihood value of 8.002 with an
associated p-value of 0.004; we can
conclude from this that the two models
are indeed different with the additional
free parameter providing a better fit.
As I mentioned before these global fit

measures are not always to be trusted for
example the following is a comparison
between our usual model and the one
where both factors load on lozenges; using
this approach we would think both models
were equally well fitted, which we know
from looking at the individual parameter
estimates is not the case
Ωnyx has many more features which I do

not have room to discuss in this chapter, for
further information visit the Ωnyx website.

65.14.2 OpenMx (formally Mx)
OpenMx is another SEM package. However this originally was a standalone application called
Mx which has been redeveloped as an R library. The OpenMx code generated by Ωnyx for
our CFA is shown below. I do not intend to discuss this, just noting that it is very different
syntactically. The OpenMx website has extensive tutorials, and because people tend to learn
either lavaan or OpenMx I will leave it at that.
# This model specification was automatically generated by Ωnyx
require("OpenMx");
modelData <- read.table(DATAFILENAME, header = TRUE)
manifests<-c("visperc","cubes","lozenges","paragrap","sentence","wordmean")
latents<-c("visual","reading")
model <- mxModel("grnt_fem_m1",
type="RAM",
manifestVars = manifests,
latentVars = latents,
mxPath(from="visual",to=c("visperc","cubes","lozenges"), free=c(TRUE,TRUE,TRUE), value=c(1.0,1.0,1.0) , arrows=1,
label=c("visual_TO_visperc","visual_TO_cubes","visual_TO_lozenges") ),
mxPath(from="reading",to=c("paragrap","sentence","wordmean"), free=c(TRUE,TRUE,TRUE), value=c(1.0,1.0,1.0) , arrows=1,
label=c("reading_TO_paragrap","reading_TO_sentence","reading_TO_wordmean") ),
mxPath(from="visperc",to=c("visperc"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_visperc") ),
mxPath(from="cubes",to=c("cubes"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_cubes") ),
mxPath(from="lozenges",to=c("lozenges"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_lozenges") ),
mxPath(from="paragrap",to=c("paragrap"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_paragrap") ),
mxPath(from="sentence",to=c("sentence"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_sentence") ),
mxPath(from="wordmean",to=c("wordmean"), free=c(TRUE), value=c(1.0) , arrows=2, label=c("VAR_wordmean") ),
mxPath(from="visual",to=c("visual"), free=c(FALSE), value=c(1.0) , arrows=2, label=c("VAR_visual") ),
mxPath(from="reading",to=c("reading","visual"), free=c(FALSE,TRUE), value=c(1.0,1.0) , arrows=2, label=c("VAR_reading","COV_reading_visual") ),
mxPath(from="one",to=c("visperc","cubes","lozenges","paragrap","sentence","wordmean"), free=T, value=1 , arrows=1),
mxData(modelData, type = "raw")
);
result <- mxRun(model)

summary(result)
65.15 Structural Equation Modelling directly in R using the sem package

From the above lavaan example, I'm sure you realise that you could have also directly
written the lavaan code in R and bypassed Ωnyx but at the same time I'm sure you can see
the advantage of using Ωnyx. An R package originally developed long before lavaan, in 2001,
called sem also allows the creation of SEM models in R directly. The sem package just uses
two types of relation: <-> to indicate variances/covariances and -> to indicate
# lavaan code regression paths. You then need to give each a name or specify it as being a fixed value
onyx.model<-"
! regressions using NA and also indicate the estimation starting/fixed value or allow the program to
visual=~visperc estimate it freely (NA). An example should help make this clearer. Because SEM model
visual=~cubes
visual=~lozenges specification code is similar to the lavaan approach I have put them side by side below.
reading=~paragrap
reading=~sentence install.packages("sem") ; library(sem) # to install and load the sem package
reading=~wordmean hozdatamodel <- specifyModel()
! residuals, variances and visual -> visperc, visual_TO_visperc, NA #first factor
covariances visual -> cubes, visual_TO_cubes, NA
visperc ~~ visperc visual -> lozenges, visual_TO_lozenges, NA
cubes ~~ cubes reading -> paragrap, reading_TO_paragrap, NA # second factor
lozenges ~~ lozenges reading -> sentence, reading_TO_sentence, NA
paragrap ~~ paragrap reading -> wordmean, reading_TO_wordmean, NA
sentence ~~ sentence visperc <-> visperc, err_visperc, NA # errors
wordmean ~~ wordmean cubes <-> cubes, err_cubes, NA
visual ~~ 1*visual lozenges <-> lozenges, err_lozenges, NA
reading ~~ 1*reading paragrap <-> paragrap, err_paragrap, NA
visual ~~ reading sentence <-> sentence, err_sentence, NA
" wordmean <-> wordmean, err_wordmean, NA
onyx.model visual <-> reading, COV_visual_reading, NA # covariances
result<-lavaan(onyx.model, visual <-> visual, NA, 1 # fixed parameters
reading <-> reading, NA, 1
data=hozdata)
summary (result,
hozmodel1 <- SEM(hozdatamodel, data = hozdata)
fit.measures= TRUE)
summary(hozmodel1)
standardizedCoefficients(hozmodel1)

There are many similarities between the two and I
would recommend anyone who is thinking about
using the sem package to begin by adapting the
lavaan output from the Ωnyx application. The SEM
package is very unhelpful when it comes to
reporting errors, providing what I find to be, very
cryptic messages. You can waste many hours only
to discover the problem was a missing comma.
The results are provided opposite to demonstrate

their equivalence to the other packages/ programs
used.
There are also similar functions in the SEM package

to produce the various residues and matrices:
effects(hozmodel1)
# gives total/direct/indirect effects
hozmodel1$C # model-reproduced covariance matrix
hozmodel1$A # RAM A matrix RAM P matrix

covariances amongst elements
same as S in Ωnyx
hozmodel1$P
residuals(hozmodel1)
standardizedResiduals(hozmodel1)
covariances amongst
vcov(hozmodel1) all parameters
I have not shown all the results for the above options as
they would produce results very similar to those shown previously.
Another useful command in the sem() function, is objective =objectiveGLS which instructs
the program to produce generalized least squares estimates instead of the default maximum
likelihood ones. This can be useful if there is a problem with the initial estimation process.
Secondly you can get the sem() function to provide details for each iteration of the
estimation process by adding the option debug=TRUE. The use of these options is shown
below.
hozmodel2 <- sem(hozdatamodel, data = hozdata, objective=objectiveGLS, debug=TRUE)
hozmodel2
same as the mi measures in
the lavaan package
The modIndices() function provides modification indices.
By setting the n.largest=n you will obtain the nth largest
values, the default is 5, as shown opposite.

The sem package has what are known as wrapper functions for the sem() function, this
basically means that these other functions set a number of defaults so that you can specify
the model by providing less information. One such wrapper function is cfa() standing for
Confirmatory Factor Analysis. I have used this function below to create an equivalent cfa
model to that produced above, called cfa_hoz, but with much less code. When you run this
code the sem package informs us that additional error variances will automatically be added
along with a covariance between the factors.
cfa_hoz <- cfa(reference.indicators=FALSE)
F1: visperc, cubes, lozenges
F2: paragrap, sentence, wordmean
hozmodel3 <- SEM(cfa_hoz, data = hozdata)
summary(hozmodel3)
I personally prefer to use the sem() function as you then know exactly what you are getting.
We have now considered both creating SEM models directly in Ωnyx, specifying them initially
in Ωnyx, then repeating the analysis in R using the automatic code generated and finally
specifying the model directly in R. I deliberately kept the example as simple as possible but
now it is time to see a more typical example to demonstrate some of the representative
characteristics of an SEM analysis.
The latest version of Ωnyx (v1.00) provides an export to the SEM package, producing the
following code on the left, as with the previous export code you need to edit it by adding the
dataframe you wish to use. Also the first call to the sem() function seems to create an error
for this particular model.
Ωnyx export Edited
require("SEM"); require("SEM");
paths <- c("Education <-> Education", "Popden <-> Popden", paths <- c("Education <-> Education", "Popden <-> Popden",
"Nonwhite <-> Nonwhite", "Mortality <-> Mortality", "Education -> "Nonwhite <-> Nonwhite", "Mortality <-> Mortality", "Education ->
Mortality", "Nonwhite -> Mortality", "Nonwhite <-> Popden", Mortality", "Nonwhite -> Mortality", "Nonwhite <-> Popden",
"Education <-> Popden", "Education <-> Nonwhite", "Popden -> "Education <-> Popden", "Education <-> Nonwhite", "Popden ->
Mortality") Mortality")
parameter <- c("VAR_Education", "VAR_Popden", "VAR_Nonwhite", parameter <- c("VAR_Education", "VAR_Popden", "VAR_Nonwhite",
"VAR_Mortality", "Education_TO_Mortality", "Nonwhite_TO_Mortality", "VAR_Mortality", "Education_TO_Mortality", "Nonwhite_TO_Mortality",
"COV_Nonwhite_Popden", "COV_Education_Popden", "COV_Nonwhite_Popden", "COV_Education_Popden",
"COV_Education_Nonwhite", "Popden_TO_Mortality") "COV_Education_Nonwhite", "Popden_TO_Mortality")
values <- c("0.7026222048487266", "2108953.5033314778", values <- c("0.7026222048487266", "2108953.5033314778",

"78.26043332764722", "1541.636507142136", "-25.50698060512576", "78.26043332764722", "1541.636507142136", "-25.50698060512576",
"3.9999225532362965", "-163.73509363742735", "-286.0402052689569", "3.9999225532362965", "-163.73509363742735", "-286.0402052689569",
"-1.5481333432299311", "0.007954486809572933") "-1.5481333432299311", "0.007954486809572933")
model <- array(c(paths, parameter, values), dim = c(10,3)) model <- array(c(paths, parameter, values), dim = c(10,3))
colnames(model) <- c("col1","col2","col3") colnames(model) <- c("col1","col2","col3")
dat <- read.table(DATAFILENAME, header = TRUE) dat <- hozdata
result <- SEM(model = model, data = dat, fixed.x ="Intercept", raw result <- SEM(model = model, data = dat)
= TRUE) summary(result)
result <- SEM(model = model, data = dat) standardizedCoefficients(result)

65.16 A complex example
Lynsky, Fergusson & Horwood, 1998, analysed the relationships between reports of tobacco,
alcohol, and cannabis use in a birth cohort of New Zealand children studied until the age of
16 (n=1265 dropping to 913 for complete cases). Comparable information from their parents
was also obtained. The
parental self parental self parental self
researchers were interested
reported reported reported reported reported reported
in assessing the validity of a
particular model known as
alcohol
cannabis
Jessor and Jessor's Problem
use
tobacco use Behaviour Theory, which
use
posits that proclivity to
vulnerability problem behaviours are due
to three aspects;
peer
affiliations parental demographic social structure
novelty use (peer afflictions, novelty
seeking
seeking and parental illicit
drug use); the perceived
Lynskey_et_al_1998_final model
social environment and the
personality system. This is in
0.22*
contrast to the stage or
0.42*
gateway theory where there
0.15*
0.18* is a progression from one
E11.0 E21.0 E31.0 E41.0 E51.0 E61.0
substance to another.
0.61* 0.34* 0.81* 0.64* 0.87* -0.51* (Lynsky, Fergusson &
Horwood, 1998 p.1004).
parental report tobacco self report tobacco parental report alcohol self report alcohol parental report cannabis self report cannabis They had therefore a general
idea of something like that
shown above.
0.79* 0.94* 0.46* 0.86*
0.59* 0.77*
The final model produced

0.73
from the study is shown
D1* 0.75 cannabis use
tobacco use
alcohol use D2* opposite. It must be
0.61
remembered that this is the
D3* final model of a long
0.79*
0.68* 0.66*
development process and I
0.15*
strongly advise you to read
0.60 the original article. The
vulnerability D4*
parameter values are slightly
different but comparable to
0.08*
0.66* 0.28*
the original as I have
parental use*
produced it using a subset of
peer affiliations* novelty seeking*
the original data (n=868) in
0.28*
0.09* the commercial SEM
program EQS.
0.10*
Figure X: EQS 6 all_correlations_normality_assumed.eds Chi Sq.=39.34 P=0.00 CFI=0.99 RMSEA=0.06

In EQS the error values are the square roots of those produced in Ωnyx.
Therefore: 1 -(SEM error value)2 = R2.
There is a lot to take in concerning this complex model. Possible key points are:
1. Parental reports of tobacco, alcohol and cannabis use are correlated in contrast to self-
reported cannabis use, which does not correlate with parental or other self reported
measures.
2. Self-reported measures of tobacco, alcohol and cannabis use have higher standardised
factor loadings (.77 to .94) compared to parental values ( .46 to .79). Remember the
higher the loading the greater the degree to which the measure is reflected in the factor.
3. Relationship between tobacco, alcohol and cannabis use (latent variable) and
vulnerability - the standardised loading factor values here (.66 to .79) indicates a strong
relationship between the three specific latent variables and the general vulnerability
latent variable.
4. The error value associated with the vulnerability (.6)

indicates how well the three predictors (peer affiliations,
novelty seeking and parental drug) predict vulnerability. As
R2 = 1-.62 = .64 indicting that 64% of the variability in
vulnerability can be accounted for by the predictors. For
the full data set (Lynsky, Fergusson & Horwood, 1998) this
value was reduced to 54%.
5. The asterisks (*) beside the parameter estimate indicates that the value is statistically
significant.
6. The overall model fit measures indicate a well-fitting model with CFI =0.99 (good fit
CFI≥0.95) and RMSEA=0.06 (RMSEA≤.05 very good fit).
65.17 Extending SEM

SEM techniques have been extended in many ways:
• Non-normal distributions
• Bayesian approaches
• Ordinal and binary data, mimicking factorial ANOVA and Logistic regression.
• Time series (also called latent growth models)
• Comparisons of multiple groups (means)
• moderation & mediation effects (see
https://www.youtube.com/watch?v=mirI5ETQRTA)
• Meta-analysis (Cheung 2015)
A good introduction to the above techniques, except moderation and mediation, can be
found in Schumacker & Lomax, 2010.
65.18 Reviewing SEM research

The following page provides a checklist, of aspects to consider when reviewing a SEM article.
While all the points are valid, one or two of them, such as cross validation of the modal using
a subset of the data is a rare occurrence.

Criteria for reviewing a article using SEM (adapted from Schumacker & Lomax 2006, p.251-5)
1. Data Preparation
1. Have you adequately described the population from which the random sample data was drawn?
2. Did you report the measurement level of your variables?
3. Did you report the descriptive statistics on your variables? [+ normality]
4. Did you test for multivariate normality?
5. Did you create a table with correlations, means, and standard deviations?
6. Did you have missing data if so why and what strategy did you use to overcome the problem.
7. Did you have outliers and did you resolve this by using robust statistics or deletion methods?
8. Did you resolve non-normality of a variable by some form of transformation?
9. Did you have problems with multi-collinearity among variables? If so how did you resolve it?
10. Did you specify the input data used (raw, correlations, covariances etc.)
11. Did you include the set of commands as an appendix so that others could carry out a similar analysis?
2. Model Specification
1. Did you provide a rationale for your study?
2. Did you explain why SEM rather than another approach was required?
3. Did you describe your latent variables
4. Did you provide a theoretical foundation for your model?
5. Did you theoretically justify alternative models?
6. Did you justify your sample size?
7. Did you clearly state your statistical hypotheses?
8. Did you discuss the expected magnitude and direction of expected parameter estimates?
9. Did you include a figure or diagram of your proposed model(s)?
3. Model Identification Issues

1. Did you calculate the number of distinct values in your sample matrix?
2. Did you calculate the number of parameters needing to be estimated (free parameters) [and constraints]
3. [Did you software provide you with appropriate messages indicating there was no identification problem.]
4. Model Estimation
1. How will you consider power and sample size?
2. What is the ratio of sample size to number of parameters?
3. What estimation technique is most suitable given your sample size and normality situation?
4. Did you encounter Heywood cases (negative variance) or other impossible values in the output?
5. Did the software use raw data or a matrix as input?
6. How did you scale the latent variable (reference variable or fix the variance of the latent variable).
7. Which SEM program and version did you use?
8. Did you encounter any convergence problems?
9. Did you report the R2 values to indicate the total effect the independent variables had on each dependent variable.
5. Model Testing
1. Did you include a website providing more information?
2. Did you provide tables (American Psychological Association style, Schumacker p.241) providing details for each factor (i.e. Reliability rho) and for each
indicator Loading and R2
3. Did you use single items or composite scale scores?
Several other issues in the original list not considered here.
6. Model Modification
1. Did you compare alternative or equivalent models?
2. Did you clearly indicate how you modified the initial model?
3. Did you provide a theoretical justification for the respecified model?
4. Did you add or delete one path/ parameter at a time? [recommended]
5. Which statistics helped guide you through the process (Wald, Lagrange [good] or simple t tests [bad])
6. Did you provide parameter estimates and model fit indices for both the initial model and the respecified one(s)?
7. Did you report expected change statistics.
8. Your model is not the only model that fits the sample data, so did you check for equivalent models or theoretically justify your final model?
9. How did you evaluate and select the best model?
7. Model Validation
1. Did you replicate your SEM model analysis using another sample of data?
2. Did you cross-validate your SEM model by splitting your original sample of data?
3. Did you use bootstrapping [simulate the samples] to determine bias in your parameter estimates?
4. Did you report effect sizes and confidence intervals in addition to statistical significance testing?
5. Did you evaluate your results with regard to your original theoretical framework?

65.19 Tricks and Tips
This chapter has provided only an introduction but hopefully has helped you understand the
basics of SEM and provided enough information to enable you to interpret findings from
such an analysis. People who undertake SEM tend to learn a particular SEM program or R
package and then stick to it. I hope by introducing a range of approaches in this chapter will
have helped you decide which is the best one for you.
Understandably, much of the information available is aimed at using a certain package,

particularly the many useful youtube videos. Notable exceptions are Loehin, 2004 and Kline,
2016. The companion website to Kline's book provides datasets and syntax files for several
chapters and an excellent set of web links to other SEM sites at:
http://www.guilford.com/cgi-bin/cartscript.cgi?page=etc/kline.html
There are many websites devoted to SEM techniques. Professor Jason Newsom, at Portland
University, runs an excellent course on SEM with online resources and links to several SEM
sites at: www.upa.pdx.edu/IOA/newsom . Another is David A. Kenny (Emeritus Professor,
University of Connecticut, Department of Psychology) who keeps a useful, succinct webpage
listing and explaining most of the model fit indices http://davidakenny.net/cm/fit.htm
There are the webpages associated with each of the R packages used in this chapter which
provide further details and examples. The Ωnyx website provides some information which
hopefully will be developed in future.
Finally I have augmented the material presented in this chapter with several online chapters
and YouTube videos, all available from the books website (http://robin-
beaumont.co.uk/rbook/sem/index.html) two of which are:
• SEM equivalents to basic statistical procedures

• Regression models in SEM

(OK) Onyx - Structural - Equation - Modelling - 2018 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(OK) Onyx - Structural - Equation - Modelling - 2018 PDF

Uploaded by

Copyright:

Available Formats

An Introduction to Structural Equation Modelling (SEM)

Robin Beaumont 02/01/2018

65.1 The background to SEM ....................................................................................................................... 2

65.1 The background to SEM

cubes 5. Extends Confirmatory Factor Analysis (CFA)

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 3

In 1980, McArdle suggested an alternative formulation, simplifying the 8 matrices needed by

Predictors of mortality While it is not necessary to

m = ( 0.00 0.00 0.00 0.00 )

• A = Asymmetrical arrows = regression

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 4

1. What is meant by difference between?

65.2.2 Interpreting a Path model (i.e. no circles)

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 5

65.2.4 Computing R squared

R 2  1 rYX1   2 rYX 2  3 rYX 3

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 6

(1 - the standardised residual variance (SRV) of Mortality) = 1-.41 = 0.59.

D 3. No more than one bi-directional arrow can be included in each path-chain.

Indirect effect of B to C = BAC admissible (rule 1)

Neurological 0.54 3. What is the direct effect of Diminished Socioeconomic

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 7

In the first the

Latent variables therefore provide a method of modelling concepts such as measurement

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 8

Related to this is the fact that authors

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 9

This ratio of the of sample size to number of parameters to be estimated in an SEM is

65.6 Constraining parameters

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 10

65.7 Model fit

1. Inspect two overall fit measures:

o RMSEA (Root Mean Square Error of Approximation) - is at present the most

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 11

A typical set of model fitting statistics is shown

The chi-squared actual value is very small (E-13)

65.7.1 RMSEA Root Mean Square Error of Approximation

1. How well does my model fit my statistical population?

65.8 Power and sample size determination

65.9 Why and When to use Structural Equation Modelling

1. Confirm Relationships / hypotheses

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 13

If you go to the link above the page

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 14

Then download Ωnyx from:

The Ωnyx program, being a java file, has a jar

As shown opposite, I created a folder on my d: drive called onyx.

65.11.2 Creating an SEM model in

The first thing you need to do is create a

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 15

A hexagonal window (called the dataset panel) appears which

Ωnyx has context-sensitive popup menus, which means that if you

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 16

We now need to add the two latent variables

Either double right mouse click in an empty part of the

Right mouse click on the model panel to bring up the

We select each latent variable (factor) in turn and then bring up

We now need to fix

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 17

Right mouse click on it and drag to the target variable.

The style described above is the traditional way of defining a factor

The far left screenshot shows that by hovering the mouse

D:\stats book_scion\new_version2016\65_structural_equation_modelling_2018.docx Book chapter 65 Page 18